Automated classification of normal and abnormal heart sounds using support vector machines

While a number of researchers have attempted to develop algorithms for automated classification of heart sounds over the last five decades, these studies have been inadequate in terms of clinical utility due to a number of important limitations. The PhysioNet/Computing in Cardiology 2016 Challenge seeks to facilitate the development of highly robust algorithms to perform automatic classification of heart sounds in a manner which overcomes the limitations of previous studies. The dataset consists of over 3000 phonocardiogram recordings, taken from several locations on the body, from both healthy and pathological adults and children. The classification task requires the algorithm to determine if a recording is normal, abnormal or cannot be scored (due to excessive noise/corruption of the signal). An implementation of a state-of-the-art segmentation algorithm has been provided by the Challenge organizers, leaving the primary focus of the Challenge on the classification task. For this task, we selected a number of features in both the time and frequency domains. For Phase I, our best overall score for the hidden test set was 0.78 (Sensitivity = 0.70, Specificity = 0.87). For Phase II, our best score was 0.7864 (Sensitivity = 0.733, Specificity = 0.8398).